KR20080045659A - Information processing device, method, and program - Google Patents

Abstract

There are provided an information processing device, an information processing method, and a program capable of suppressing concentration of recommendations in some contents and recommending a content even to a user having little history information by the CF method. Step 11 detects a user A for which a music composition is recommended and another user X having most similar history information. Step 12 detects a music composition owned by the user X and not owned by the user A. Step 13 identifies a cluster of each cluster layer to which the music composition "a" belongs. Step 14 extracts a music composition classified to be common to all the identified clusters as a recommended candidate. Furthermore, step S15 selects one music composition having cluster information most similar to the music composition "a" among the recommended candidates. The music composition thus selected is recommended to the user A. The present invention may be applied, for example, to a content sale site established in the Internet.

Description

Information processing apparatus, methods, and programs {INFORMATION PROCESSING DEVICE, METHOD, AND PROGRAM}

BACKGROUND OF THE INVENTION Field of the Invention The present invention relates to an information processing apparatus, an information processing method, and a program, and more particularly, to classify content into clusters, to manage the characteristics of the content by using the cluster in which the content is classified, and to use the content for search and recommendation. An information processing apparatus, an information processing method, and a program are related.

Background Art Conventionally, an invention for retrieving and recommending content such as a television program or a piece of music based on a user's preference (so-called content personalization) has been proposed (see Patent Document 1, for example).

In content personalization, a method called cooperative filtering (CF) or a method called content-based filtering (CBF) is widely used.

The CF method manages a purchase history of each user, detects another user X having a similar purchase history for user A who wants to recommend content, and the other user X purchases, and user A does not purchase. In order to recommend content that is not used, for example, it is employed in a mail order site on the Internet.

In the CBF method, metadata previously provided by the distribution side or sales side with respect to the content has been directly used for extracting preferences and recommending content. That is, the distance (cosine correlation or the like) between the feature vector representing the user's preference and the feature vector of each of the candidate music pieces is calculated, and the short-distance music calculated is recommended as the user's preference.

[Patent Document 1] Japanese Patent Application Laid-Open No. 2004-194107

<Start of invention>

Problems to be Solved by the Invention

The following problems arise from the above-described CF method.

(1) Content not purchased by any user is not recommended to anyone. Therefore, among the contents prepared in a large scale, recommended contents concentrate on some parts, and the majority of the contents are not recommended.

(2) In the case of recommending content to a new user, since the purchase history of the user is small, other users with similar history information cannot be detected and the content cannot be recommended (so-called cold start problem).

(3) In general, the number of users of the content also increases, but in that case, the operation when detecting other users having similar purchase histories increases, so that the recommended content cannot be determined quickly.

The present invention has been made in view of such a situation, and the CF method prevents the recommendation from being concentrated on some of the contents, and also allows the content to be recommended even for a user with little history information.

Means for solving the problem

An information processing apparatus, which is an aspect of the present invention, is an information processing apparatus that selects a content satisfying a predetermined condition from a content group and presents the same to a user, wherein each content constituting the content group is selected according to the metadata of the content. Holding means for holding a content classification means for classifying any one of the plurality of first clusters in each of the hierarchies, and a database indicating a correspondence relationship between each content and the first cluster in the hierarchy where each of the respective content is classified; Management means for managing history information on the content of the user, selection means for specifying a first cluster of interest based on the history information, and selecting content classified in the specific first cluster; Presentation means for presenting the selected content.

The selection means includes detection means for detecting a second user having a similar history information with the first user, and content not present on the history information of the first user and present on the history information of the second user. Specifying means for specifying a classified first cluster, and extracting means for extracting content classified in the identified first cluster, wherein the presenting means presents the extracted content for the first user. You can do that.

An information processing apparatus which is an aspect of the present invention includes generation means for generating preference information indicating a preference of the user in units of the first cluster based on history information of the user and the database, and a user based on the preference information. Grouping means for grouping the second means, wherein the selecting means comprises: detecting means for detecting a second user belonging to the same group as the first user, and not present on the history information of the first user; Specifying means for specifying a first cluster in which the content existing on the user's history information is classified, and extracting means for extracting the content classified in the specified first cluster, wherein the presenting means comprises: the first means; 1, the extracted content can be presented to the user.

The information processing apparatus which is an aspect of the present invention further includes generation means for generating preference information indicating the preference of the user in the first cluster unit based on the history information of the user and the database, and in the selection means, Detecting means for detecting a second user having a similar symbol represented by the first user and the preference information, and specifying a first cluster to be noted based on the preference information of the first user and preference information of the second user. A specific means and extraction means for extracting content classified in the specified first cluster are provided, and the presentation means can present the extracted content for the first user.

The detecting means includes normalization means for normalizing user's preference information, weight calculation means for calculating weight for each layer for each of the users from preference information of each normalized user, weight for each hierarchy and the preference information. A similarity calculating means for calculating a similarity indicating a degree of similarity between the first user among the users and the other user among the users is provided, and from the calculated similarity, a second user having similar tastes with the first user can be detected. Can be.

An information processing apparatus, which is an aspect of the present invention, includes generation means for generating preference information indicating a preference of the user in units of the first cluster based on history information of the user and the database, and based on the preference information. And grouping means for grouping the users, wherein the selection means includes: detection means for detecting a second user belonging to the same group as the first user, preference information of the first user, and preference information of the second user. Specifying means for specifying a first cluster to be noted based on the above, and extracting means for extracting content classified in the identified first cluster, and the presenting means includes the content extracted for the first user. Can be presented.

In the information processing apparatus which is an aspect of the present invention, the setting means for setting a keyword for each of the first clusters in which the metadata is classified by the content classification means, and the keyword set by the setting means are used. Creation means for creating a reason statement indicating the reason for presentation is further provided, and the reason statement can also be presented to the presentation means.

Metadata classification means for classifying the metadata of the content into any one of the plurality of second clusters, and assigning the hierarchy to the second cluster is further provided, and the content classification means includes each content of the assigned hierarchy. Each can be classified into any one of a plurality of first clusters.

Generation means for generating preference information indicating the preference of the user in units of the first cluster based on the history information of the user and the database, wherein the selection means includes the first of all the hierarchies of all the hierarchies; Among the clusters, the content classified into the largest number of the first clusters can be selected as the first cluster represented by the preference information.

An information processing method, which is an aspect of the present invention, is an information processing method of an information processing apparatus for selecting and presenting content that satisfies a predetermined condition from a content group to a user. Classify the data into any one of a plurality of clusters in each hierarchy according to metadata, and maintain a database indicating a correspondence relationship between each content and each cluster in the hierarchy in which each content is classified, and Managing history information, specifying a cluster of interest based on the history information, selecting content classified in the specific cluster, and presenting the selected content.

A program, which is an aspect of the present invention, is a program for selecting content that satisfies a predetermined condition from a content group and recommending it to a user, wherein each content constituting the content group is included in each layer according to the metadata of the content. Classify the content into one of a plurality of clusters, maintain a database indicating a correspondence relationship between each content and the cluster in the hierarchy in which each content is classified, and manage history information on the content of the user; Based on the history information, the computer executes a process including the step of specifying a cluster of interest, selecting content classified in the specific cluster, and presenting the selected content.

In one aspect of the present invention, each content constituting the content group is classified into any one of a plurality of clusters in each hierarchy according to the metadata of the content. Then, a database showing a correspondence relationship between each content and the cluster in the hierarchy in which each content is classified is maintained. Further, history information on the user's content is managed. Based on the history information, the cluster of interest is specified, the content classified in the specified cluster is selected, and the selected content is presented.

Effect of the Invention

As described above, according to one aspect of the present invention, it becomes possible to recommend content to a user by the CF method.

In addition, according to one aspect of the present invention, it becomes possible to suppress the concentration of the recommendation to a part of the content of the entire content.

Further, according to one aspect of the present invention, it is possible to recommend content even to a user who has little history information.

1 is a block diagram showing a configuration example of a recommendation system to which the present invention is applied.

2 is a diagram showing the concept of clusters and cluster layers for classifying metadata of music pieces;

3 is a diagram showing an example of a music-cluster correspondence table;

4 is a diagram showing an example of a cluster-music correspondence table.

5 is a diagram illustrating an example of a user's preference vector.

6 is a flowchart for explaining preprocessing when offline.

7 is a flowchart for explaining a first recommendation process.

8 is a flowchart for explaining second and third recommendation processes.

9 is a flowchart for explaining a fourth recommendation process.

10 is a flowchart for describing fifth and sixth recommendation processes.

11 is a flowchart for explaining a seventh recommendation process.

12 is a block diagram illustrating a configuration example of a general-purpose personal computer.

It is a block diagram which shows the example of another structure of the recommendation system of one Embodiment of this invention.

14 is a flowchart for explaining another example of preprocessing when offline.

Fig. 15 is a diagram showing an example of metadata of each piece of music soft-clustered.

16 is a diagram illustrating an example of metadata of each piece of music.

Fig. 17 is a diagram showing an example of metadata of each clustered music.

18 is a block diagram illustrating an example of a configuration of a similar user detection unit.

Fig. 19 is a flowchart for explaining processing of detection of user X having similar preferences.

20 is a diagram illustrating an example of a symbol vector.

21 illustrates an example of a normalized symbol vector.

22 is a diagram illustrating an example of weights.

Fig. 23 is a diagram showing an example of similarity calculated without weighting.

24 is a diagram illustrating an example of similarity calculated by weighting.

<Description of the code>

1: Recommended system

11: Music DB

12: clustering part

13: Keyword setting part

14: clustering completed DB

15: cluster-music correspondence list

16: Music piece-cluster correspondence list

17: User history information DB

18: Recommendation candidate selection part

19: symbol vector generator

20: Similar user detection unit

21: user grouping unit

22: difference detection unit

23: Recommended cluster decision part

24: extraction part

25: music selection part

26: Newness judgment department

27: selection reason generator

28: presentation part

100: personal computer

101: CPU

111: recording medium

201: metadata clustering unit

202: music clustering part

203: Similar user detection unit

231: normalization

232: weight calculation unit

233: Similarity calculator

EMBODIMENT OF THE INVENTION Hereinafter, the specific embodiment which applied this invention is described in detail, referring drawings.

1 shows a configuration example of a recommendation system according to an embodiment of the present invention. This recommendation system 1 manages the user's history information (information such as the purchase, viewing, retrieval, retention, etc. of music data) at the sales site of music data established on the Internet, for example, and uses the CF method. Select the recommended piece of music and present it to the user. In addition, the recommendation system 1 can be applied to sales sites that sell contents other than music, for example, television programs, movies, books, and the like.

The recommendation system 1 is provided in a music database (DB) 11 and a music database 11 in which metadata of numerous pieces of music data (hereinafter, simply referred to as music) for recommending and selling to users is recorded. On the basis of the recorded metadata of each piece of music, a clustering unit 12 for clustering each piece of music to generate cluster information for each piece of music, and for setting each of the cluster layers and keywords representing the characteristics of each cluster in the cluster layer, respectively. It consists of a keyword setting part 13 and a clustering completed database (DB) 14 which holds the clustering result of each music.

In the clustering completion DB 14, as a clustering result, a cluster-music correspondence table 15 indicating a piece of music belonging to each cluster and a music-cluster correspondence table 16 indicating a cluster to which each piece of music belongs are held.

In addition, the recommendation system 1 includes a recommendation candidate selection unit 18 that selects a plurality of pieces of music to be recommended candidates based on the user history information database (DB) 17 that manages the history information of each user, and the user information. ), A music selection section 25 for selecting one piece of music from a plurality of selected recommendation candidates, a novelty determination section 26 for determining whether or not the selected piece of music is novel to the user to be recommended, and the selected piece of music to the user. The selection reason generation part 27 which produces | generates the recommendation reason statement at the time of presentation, and the presentation part 28 which presents the selected piece of music and the recommendation reason statement to a user are comprised.

The recommendation candidate selection unit 18 includes a preference vector generation unit 19, a user grouping unit 20, a similar user detection unit 21, a difference detection unit 22, a recommendation cluster determination unit 23, and an extraction unit 24. ).

The music DB 11 maintains metadata of music recommended and sold like CDDB (CD Data Base) or Music Navi, which is a data server on the Internet, which supplies metadata of music recorded on a music CD.

The clustering unit 12 includes all items (artist name, genre, album, artist review, music review, title, tempo, beat, rhythm, etc.) of the metadata of the music for all the music of the music DB 11 or their music. Based on the combination (tempo, beat, rhythm, etc.), cluster layers (first to n layers) as shown in FIG. 2 are made, and one or more of a plurality of clusters provided with music in each cluster layer. Classify (cluster).

Although music is described as an example here, artists and albums are similarly clustered in multiple layers using a lot of meta. For music recommendation, artist recommendation, and album recommendation, multilayer music clusters for artists, multilayer clusters for artists, and multilayer clusters for albums are used, respectively.

Clustering may be used in any manner, but the clustering method and distance measure are selected for each cluster layer. For example, if the actual information of the metadata is a numerical attribute such as tempo, and if it is a title attribute such as a title, it is numerically determined using a quantification method such as principal component analysis to define and cluster distance measures such as Euclidean distance. do. Representative clustering methods include K-means method, hierarchical clustering method (group average method, longest distance method, word method), soft clustering method and the like.

At this time, it is preferable to carry out by clustering (for example, clustering with restrictions) which reflected the symbol distance. For this purpose, partial surveys (symbolic sets of near real information, sets of far real information, etc.) are made by preliminary investigation, and appropriate numerical expressions, distances, and clustering methods are used. In addition, it is preferable to select a clustering method (that is, a clustering method having different characteristics) in which the independence of each cluster layer to be formed is increased.

Further, one piece of real information may be classified into a plurality of clusters of the same cluster layer. The distance (showing similarity) between clusters existing in the same cluster layer is assumed to be known. This clustering method will be described later. Instead of metadata, cluster information including cluster IDs (CL11 in FIG. 2, etc.) of clusters in which real information of each item of metadata is classified as information representing the characteristics of a piece of music is generated and the clustering completed DB 14 Output to

In addition, when a cluster suitable for classification does not exist, you may newly establish a cluster. The size of each cluster can optionally contain a plurality of real information. In addition, a cluster capable of classifying only single real information may be provided. In this case, the ID (artist ID, album ID, title ID) of the real information that can be uniquely classified may be used as the cluster ID of the cluster.

The clustering completion DB 14 generates and maintains the cluster-music correspondence table 15 and the music-cluster correspondence table 16 based on the cluster information of each piece of music generated by the clustering unit 12. In addition, the clustering completion DB 14 also maintains the keywords set for each cluster layer and each cluster set by the keyword setting unit 13.

3 shows an example of the music-cluster correspondence table 16. In FIG. 3, cluster information of the music of music ID = ABC123, for example shows that it is (CL12, CL21, CL35, CL47, CL52, ..., CLn2). For example, the cluster information of the music of music ID = CTH863 shows that it is cluster ID (CL11, CL25, CL31, CL42, CL53, ..., CLn1).

FIG. 4 shows an example of the cluster-music correspondence table 15 according to the music-cluster correspondence table 16 shown in FIG. In FIG. 4, for example, music ID = CTH863 corresponds to cluster ID = CL11. For example, it shows that cluster ID = CL21 corresponds to music ID = ABC123. For example, it is shown that cluster ID = CL32 corresponds to music ID = XYZ567.

In addition, the process of the clustering part 12, the keyword setting part 13, and the clustering completion DB 14 needs to be performed whenever the metadata of a new music is added to the music composition DB11.

Return to FIG. In the user history information DB 17, the history information which shows the music which each user purchased, watched or searched at the said sales site, or the music which the user purchased and reported that it has already possessed is hold | maintained. In the user history information DB 17, a symbol vector of each user generated by the symbol vector generation unit 19 is held. In the user history information DB 17, the grouping result of the user by the user grouping unit 20, that is, information indicating which user group each user belongs to is held.

The symbol vector generation unit 19 generates a multidimensional symbol vector having all clusters as one dimension for each user based on the history information of each user held in the user history information DB 17, It outputs to the user history information DB17. Specifically, for each piece of music in the user's history information, the music-cluster correspondence table 16 of the clustered DB 14 is referred to, and a predetermined value is added to the dimension of the symbol vector corresponding to the cluster to which the piece belongs. do. The generated preference vector of each user is managed in the user history information DB 17. When the user's history information is updated by purchasing music, etc., the symbol vector is also updated.

For convenience of explanation, as shown in FIG. 5, all the clusters are shown in FIG. 5. In the first layer, three clusters CL11, CL12 and CL13 are used. In the second layer, four clusters CL21, CL22, CL23 and CL24 are used. It is assumed that only three clusters CL41, CL42, CL43 exist in the three clusters CL31, CL32, CL33 and the fourth layer. In this case, the symbol vector is 13 dimensions.

For example, if there are records in which two pieces of music are purchased in the history information of the user A, and the cluster IDs to which the first piece of music belongs are CL11, CL22, CL33, and CL41, 1 is added to the values of the corresponding dimensions. If the cluster ID to which the second piece of music belongs is CL12, CL24, CL32, CL43, 1 is added to the values of the dimensions corresponding to them. Then, symbol A (1, 1, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 1) of user A is generated.

For example, if there are records in which three pieces of music are purchased in the history information of the user X, and the cluster IDs to which the first piece of music belongs are CL11, CL22, CL32, and CL43, 1 is added to the values of the corresponding dimensions. If the cluster ID to which the second piece of music belongs is CL12, CL22, CL33, CL42, 1 is added to the value of the dimension corresponding to them, respectively. Further, if the cluster ID to which the third piece of music belongs is CL13, CL24, CL33, CL41, 1 is added to the value of the dimension corresponding to them, respectively. The symbol vector 1, 1, 1, 0, 2, 0, 1, 0, 1, 2, 1, 1, 1 of the user X is generated.

In addition, the addition value to each dimension may be changed according to the kind of history information (that is, purchase, viewing, searching, or holding). For example, you may add 1 if it is a purchase or retention, 0.5 if it is a city hall, and 0.3 if it is a search.

Return to FIG. The user grouping unit 20 groups all the users based on the similarity of the preference vector of each user held in the user history information DB 17. However, in order to facilitate the process of determining the similarity of the multidimensional symbol vectors, the values of the respective dimensions of the symbol vector of each user are binarized. That is, when the value of each dimension is 1 or more, it is substituted by 1, and when it is 0, it is set to 0. By binarizing the values of the respective dimensions of the symbol vector in this manner, it is possible to reduce the amount of computation required for the determination of similarity as compared with the case of not binarizing, so that grouping can be easily performed. This grouping result is managed by the user history information DB17.

The similar user detection unit 21 compares the history information of each user held in the user history information DB 17 to detect a user whose music is recommended and another user whose history information is similar. In addition, the similar user detection unit 21 compares the preference vector of each user held in the user history information DB 17 to detect another user whose music is recommended and similar in preference. The difference detection section 22, based on the history information of each user held in the user history information DB 17, history information of the user whose music is recommended and the other users detected by the similar user detection section 21. To detect the difference. The recommendation cluster determination unit 23 determines the recommendation cluster based on the difference between the preference vector and the user whose music is recommended and the other user detected by the similar user detection unit 21. The extraction part 24 extracts the music used as a recommendation candidate based on the processing result of the similar user detection part 21-recommendation cluster determination part 23, and the clustering completed DB14.

The music selection unit 25 selects one piece of music from a plurality of extracted pieces of music according to a predetermined condition. For example, one belonging to more recommendation clusters, one belonging to a recommendation cluster in a predetermined high priority cluster layer, or one piece of music is randomly selected, and the selection result is determined by the novelty determination unit 26. , And the selection reason generator 27 is output. The novelty determination unit 26, based on the preference vector of the user whose music is recommended, when the degree of redundancy of the cluster to which the selected music belongs belongs is higher than or equal to a predetermined ratio (for example, 30%), the novelty is determined. If it is determined that there is no, and it is less than the predetermined ratio, it is determined that there is novelty, and the determination result is output to the selection reason generation unit 27.

The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Also, based on the determination result from the novelty determination unit 26, for example, the reason for the selection is to include the words "unexpected" for the novelty and "always" or "friendly" for the absence of the novelty. Create a statement. The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music.

Further, the review text of the selected piece of music may be cited as a selection reason statement as it is, or the selection reason statement may be generated using a word extracted from the review text of the selected piece of music. In addition, the Tf / idf method may be applied to extract a word used for the selection reason text from the review text.

The presentation unit 28 obtains information on the selected piece of music from the piece of music DB and presents it to the user along with the generated selection reason statement.

Next, the operation of the recommendation system 1 will be described. First, the offline preprocessing in preparation for a process for recommending a piece of music will be described with reference to the flowchart of FIG. 6.

In step S1, the clustering unit 12 classifies all pieces of music of the music DB 11 into any one of the item cluster layers (first to nth layers) of the metadata of the piece of music, and classifies actual information of each item. Classification (clustering) is carried out in any one of the some cluster provided in the classification cluster layer. Instead of the metadata, the clustering unit 12 generates cluster information consisting of cluster IDs of clusters in which real information of each item of metadata is classified, and generates information on the clustering completed DB 14 instead of metadata. Output In addition, clustering may be abbreviate | omitted about the music which has already been clustered, and you may make it cluster only about the music which has not been clustered. The clustering completion DB 14 generates the cluster-music correspondence table 15 and the music-cluster correspondence table 16 based on the cluster information of each piece of music generated by the clustering unit 12.

In step S2, the symbol vector generation unit 19 of the recommendation candidate selection unit 18 generates a symbol vector for each user based on the history information of each user held in the user history information DB 17. And output to the user history information DB 17. In step S3, the user grouping unit 20 groups all the users based on the similarity of the preference vector of each user held in the user history information DB 17. FIG. However, in order to facilitate the process of determining the similarity of the multidimensional symbol vectors, the values of the respective dimensions of the symbol vector of each user are binarized. The grouping result is then output to the user history information DB 17. The preprocessing at the time of offline is complete | finished above.

In this manner, the clustering of all pieces of music held in the music DB 11, generation of the preference vector of each user, and grouping of users as preprocessing can quickly execute the first to seventh recommended processes described later. have. In addition, since some user information is not used among the 1st-7th recommendation processes, when only the recommendation process which does not use the user's group information is performed, the process of step S3 may be abbreviate | omitted.

Next, a 1st recommendation process is demonstrated with reference to the flowchart of FIG. Hereinafter, the user whose music is recommended is described as user A. FIG. This process is started, for example, when the user A accesses the sales site.

In step S11, the similar user detection unit 21 compares the history information of the user A held in the user history information DB 17 with the history information of another user, whereby the other user X whose user information is the most similar in history information. Detect. In step S12, the difference detection unit 22 is based on the history information of the user A and the user X held in the user history information DB 17, which the user X has (purchased or held in the past). As a piece of music, a piece of music that the user A does not have is detected. When a plurality of pieces of music meeting this condition exist, one piece of music is selected at random. The detected music is referred to as music a.

In step S13, the recommended cluster determination unit 23 refers to the music-cluster correspondence table 16 of the clustering completed DB 14 and specifies the cluster of each cluster layer to which the music a belongs. In step S14, the extraction part 24 refers to the cluster-music correspondence table 15 of the clustering completed DB14, and extracts the music classified in common to all the clusters specified by the process of step S13. The music extracted here is used as a recommendation candidate. There may be a plurality of recommendation candidates. In addition, when there is no music classified in common in all the specific clusters by the process of step S13, among the specific clusters in the process of step S13, the music classified in common to as many clusters as possible can be extracted and recommended. It is a candidate.

In step S15, the music selection unit 25 selects one piece of music of which the candidate candidate most similar to the piece of music a detected in step S12 is clustered among the pieces of the recommendation candidates, and the selection result is determined by the novelty determination unit 26 and the reason for selection. Output to the generation unit 27. In step S16, the novelty determination unit 26 determines the presence or absence of novelty based on the preference vector of the user A and the cluster to which the selected music belongs, and outputs the determination result to the selection reason generation unit 27. The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Further, the selection reason statement is generated also on the basis of the determination result from the novelty determination unit 26.

The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music. In step S17, the presentation unit 28 obtains the information about the selected piece of music from the piece of music DB and presents it to the user along with the generated selection reason statement. This is the end of the first recommendation process.

Next, the 2nd and 3rd recommendation process is demonstrated with reference to the flowchart of FIG. First, the second recommendation process will be described. This process is started, for example, when the user A accesses the sales site.

In step S21, the similar user detection unit 21 compares the user's A's preference vector held in the user history information DB 17 with the preference vector of another user, thereby selecting another user X having the most similar user A and the preference vector. Detect. In addition, the similarity between the symbol vector of user A and the symbol vector of another user is judged by calculating cosine correlation values of both, for example.

In step S22, the difference detection unit 22 detects the dimension of the symbol vector whose value is 0 in the symbol vector of the user A and a value other than 0 in the symbol vector of the user X, and recommends a cluster corresponding to the detected dimension. Decide on

For example, the symbol vector of user A is (1, 1, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 1) shown in A of FIG. When the vector is (1, 1, 1, 0, 2, 0, 1, 0, 1, 2, 1, 1, 1) shown in B of FIG. 5, as shown by the diagonal line in B of FIG. , Clusters CL13, and CL42 are determined as recommended clusters.

In step S23, the extraction unit 24 refers to the cluster-music correspondence table 15 of the user history information DB 17 and the clustering completed DB 14, and among the pieces of music classified in the recommended cluster, the user X is selected. It extracts what the user A does not have and makes it a recommendation candidate.

In step S24, the music selection unit 25 selects one piece of music from the recommendation candidates by any one of the following three methods or a combination of three methods, and selects the result of the novelty determination unit 26. And the selection reason generation unit 27. The first method is to select pieces of music that belong to more recommended clusters. In the second method, priorities are given to the cluster layer in advance, and music pieces classified in the recommended clusters belonging to the higher priority cluster layer are selected. The third method is a method of randomly selecting.

In step S25, the novelty determination unit 26 determines the presence or absence of novelty based on the preference vector of the user A and the cluster to which the selected music belongs, and outputs the determination result to the selection reason generation unit 27. The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Also, the selection reason statement is generated based on the determination result from the novelty determination unit 26. The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music. In step S26, the presentation part 28 acquires the information about the selected music from the music DB, and presents it to the user with the generated selection reason statement. This is the end of the second recommendation process.

Next, the third recommendation process will be described. In the process of step S23 in the above-described second recommendation process, the third recommendation process is to extract and extract as a recommendation candidate the user A does not have among all the music classified in the recommendation cluster. In other words, it is also possible to make a recommendation candidate that the user X does not have. Since other processes are the same as those in the second recommendation process, description thereof is omitted.

Next, the 4th recommendation process is demonstrated with reference to the flowchart of FIG. This process is started, for example, when the user A accesses the sales site.

In step S41, the similar user detection unit 21 randomly determines another user X belonging to the same group as the user A based on the group information of the user A held in the user history information DB 17. In step S42, the difference detection part 22 is a music which user X has based on the history information of user A and user X hold | maintained in the user history information DB 17, and the music which user A does not have is contained. Detect. When a plurality of pieces of music meeting this condition exist, one piece of music is selected at random. The detected music is referred to as music a.

In step S43, the recommended cluster determination unit 23 refers to the music-cluster correspondence table 16 of the clustering completed DB 14 and specifies the cluster of each cluster layer to which the music a belongs.

In step S44, the extraction part 24 refers to the cluster-music correspondence table 15 of the clustering completion DB14, and extracts the music classified in common to all the clusters specified by the process of step S43. The music extracted here is used as a recommendation candidate. There may be a plurality of recommendation candidates. In addition, when there are no music classified in common in all the specific clusters by the process of step S43, among the specific clusters by the process of step S43, the music classified in common to as many clusters as possible can be extracted and recommended. It is a candidate.

In step S45, the music selection unit 25 selects one piece of music of which the candidate candidate most similar to the piece of music a detected in step S42 among cluster pieces of recommended candidates, and selects the result of the novelty determination unit 26 and the reason for selection. Output to the generation unit 27. In step S46, the novelty determination unit 26 determines the presence or absence of novelty based on the preference vector of the user A and the cluster to which the selected music belongs, and outputs the determination result to the selection reason generation unit 27. The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Further, the selection reason statement is generated also on the basis of the determination result from the novelty determination unit 26.

The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music. In step S47, the presentation unit 28 obtains information on the selected piece of music from the piece of music DB and presents it to the user along with the generated selection reason statement. This is the end of the fourth recommendation process.

In the fourth recommendation process, since the group information of the users grouped by the preprocessing during off-line is used, the user X similar to the history of the user A can be quickly determined.

Next, the 5th and 6th recommendation processes are demonstrated with reference to the flowchart of FIG.

First, the fifth recommendation process will be described. This process is started, for example, when the user A accesses the sales site.

In step S51, the similar user detection unit 21 randomly determines another user X belonging to the same group as the user A based on the group information of the user A held in the user history information DB 17. FIG.

In step S52, the difference detection unit 22 detects the dimension of the symbol vector whose value is 0 in the symbol vector of the user A and a value other than 0 in the symbol vector of the user X, and recommends a cluster corresponding to the detected dimension. Decide on

In step S53, the extraction unit 24 refers to the cluster-music correspondence table 15 of the user history information DB 17 and the clustering completed DB 14, and among the pieces of music classified in the recommended cluster, the user X is selected. It extracts what the user A does not have and makes it a recommendation candidate.

In step S54, the music selection unit 25 selects one piece of music from the recommendation candidate by any one of the following three methods or a combination of three methods, and selects the result of the novelty determination unit 26. And the selection reason generation unit 27. The first method is to select pieces of music that belong to more recommended clusters. In the second method, priorities are given to the cluster layer in advance, and music pieces classified in the recommended clusters belonging to the higher priority cluster layer are selected. The third method is a method of randomly selecting.

In step S55, the novelty determination unit 26 determines the presence or absence of novelty based on the preference vector of the user A and the cluster to which the selected music belongs, and outputs the determination result to the selection reason generation unit 27. The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Further, the selection reason statement is generated also on the basis of the determination result from the novelty determination unit 26. The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music. In step S56, the presentation unit 28 obtains the information about the selected piece of music from the piece of music DB, and presents it to the user along with the generated selection reason statement. This is the end of the fifth recommendation process.

Next, a sixth recommendation process will be described. In the sixth recommendation process, in the processing of step S53 in the fifth recommendation process described above, among the pieces of music classified in the recommendation cluster, the user A does not have a feature to be extracted to be a recommendation candidate. In other words, it is also possible to make a recommendation candidate that the user X does not have. Since other processes are the same as those in the fifth recommendation process, the description thereof is omitted.

In the fifth and sixth recommendation processes, since the group information of the users grouped by the preprocessing during off-line is used, the user X similar to the history of user A can be quickly determined.

Next, a seventh recommendation process will be described with reference to the flowchart in FIG. 11. First, the seventh recommendation process will be described. This process is suitable for the case where the user A's history information is extremely small, when there are few other users, and the like, for example, when the user A accesses the sales site.

In step S61, the difference detection part 22 detects that the value is more than a predetermined value among each dimension of the symbol vector of the user A, and determines the cluster corresponding to the dimension as a recommended cluster.

In step S62, the extraction unit 24 refers to the cluster-music correspondence table 15 of the user history information DB 17 and the clustering completed DB 14, and among the pieces of music classified in the recommended cluster, the user A What they do not have is extracted to be a candidate for recommendation.

In step S63, the music selection unit 25 selects one piece of music belonging to the most recommended cluster among the pieces of the recommendation candidates, and outputs the selection result to the novelty determination unit 26 and the selection reason generation unit 27. do. When a plurality of pieces of music belonging to the most recommended clusters exist, one piece of music is randomly selected, for example.

In step S64, the novelty determination unit 26 determines the presence or absence of novelty based on the preference vector of the user A and the cluster belonging to the selected music, and outputs the determination result to the selection reason generation unit 27. The selection reason generation unit 27 obtains from the clustering completed DB 14 the cluster layer to which the selected music belongs or the keyword corresponding to the cluster, and generates a selection reason statement indicating the reason for selection by using the acquired keyword or the like. Further, the selection reason statement is generated also on the basis of the determination result from the novelty determination unit 26. The generated selection reason statement is output to the presentation unit 28 together with the music ID of the selected music. In step S65, the presentation unit 28 obtains the information about the selected piece of music from the piece of music DB and presents it to the user along with the generated selection reason statement. This is the end of the seventh recommendation process.

According to the above-mentioned recommendation processing of the first to the seventh, since the CF method is applied by replacing the user's history information with a symbol vector having each cluster as one-dimensional, the recommendation of the music to the user A is the music DB 11 It can be forbidden to focus on some of the pieces of music present in. In addition, a piece of music can be recommended even for a user with little history information, so that a so-called cold start problem can be avoided. In addition, the reason why the recommended piece of music was selected can be presented to the user A. For example, the user A can know whether the recommended piece of music is novel to himself.

In addition, the present invention can be applied not only to recommending a piece of music but also to a sales site that sells contents other than the piece of music, such as a television program, a movie, a book, and the like.

By the way, although the above-mentioned series of processes can be performed by hardware, it can also be performed by software. When a series of processes are performed by software, the program which comprises the software can perform various functions by installing the computer or the various programs which are integrated in the dedicated hardware, for example, FIG. It is installed in a general-purpose personal computer configured as shown in 12 from a recording medium.

This personal computer 100 incorporates a CPU (Central Processing Unit) 101. The input / output interface 105 is connected to the CPU 101 via the bus 104. The bus 104 is connected to a ROM (Read Only Memory) 102 and a RAM (Random Access Memory) 103.

The input / output interface 105 includes an input unit 106 composed of an input device such as a keyboard or a mouse for inputting an operation command by a user, and a display such as a cathode ray tube (CRT) or liquid crystal display (LCD) for displaying a screen. The communication unit comprises an output unit 107, a storage unit 108 including a hard disk drive for storing programs and various data, and a modem, a local area network (LAN) adapter, or the like, for communication processing via a network represented by the Internet. The communication unit 109 to be executed is connected. Also includes magnetic discs (including flexible discs), optical discs (including Compact Disc-Read Only Memory (CD-ROM), digital versatile discs (DVD)), magneto-optical discs (Mini Disc) Or a drive 110 that reads and writes data to a recording medium 111 such as a semiconductor memory.

The program causing the personal computer 100 to execute the above-described series of processes is supplied to the personal computer 100 in a state stored in the recording medium 111, read by the drive 110, and stored in the storage unit 108. It is installed on the internal hard disk drive. The program installed in the storage unit 108 is loaded into the RAM 103 from the storage unit 108 and executed by a command of the CPU 101 corresponding to a command from the user input to the input unit 106. .

FIG. 13 is a block diagram showing an example of another configuration of the recommendation system 1 according to the embodiment of the present invention. In Fig. 13, the same parts as those shown in Fig. 1 are denoted by the same reference numerals, and the description thereof is omitted.

The recommendation system 1 shown in FIG. 13 includes a music DB 11, a keyword setting unit 13, a clustering completed DB 14, a user history information DB 17, a recommendation candidate selection unit 18, and a music selection. A unit 25, a novelty determination unit 26, a selection reason generation unit 27, a presentation unit 28, a metadata clustering unit 201, and a music clustering unit 202 are provided.

The metadata clustering unit 201 clusters the metadata of each piece of music recorded in the piece music database 11. In other words, the metadata clustering unit 201 classifies the metadata of the piece of music as the content into any one of the plurality of clusters and assigns a cluster hierarchy.

The metadata clustering unit 201 supplies the music clustering unit 202 with the result of clustering metadata of each music.

The music clustering unit 202 generates cluster information of each piece of music by clustering the pieces of music in the same manner as the clustering unit 12 based on the clustering result of metadata of each piece of music by the metadata clustering unit 201. do. That is, the music clustering unit 202 generates cluster information according to the clustering result of each music and outputs it to the clustering completed DB 14.

In addition, the recommendation candidate selection unit 18 of the recommendation system 1 shown in FIG. 13 includes a preference vector generation unit 19, a user grouping unit 20, a difference detection unit 22, and a recommendation cluster determination unit 23. , Extraction section 24, and similar user detection section 203.

The similar user detection unit 203 compares the preference vector of each user held in the user history information DB 17 to detect a user whose music is recommended and another user whose similar preference vector is similar. More specifically, the similar user detection unit 203 normalizes a symbol vector, which is an example of user's preference information, calculates a weight for each user for each user from the normalized user's preference vector, and calculates the weight for each user. Similarity indicating the degree of similarity between symbols between users is calculated from the weight vector and the symbol vector, and from the calculated similarity, a second user having a similar preference to the first user is detected.

Next, with reference to the flowchart of FIG. 14, another example of the preprocessing at the off-line which is preparation of the process which recommends a music in the recommendation system 1 shown in FIG. 13 is demonstrated.

In step S201, the metadata clustering unit 201 acquires the metadata of the music from the music DB 11 and compresses the dimension of the acquired metadata. For example, in step S201, the metadata clustering unit 201 determines the dimensions of the metadata of the music obtained from the music DB 11 by the latent semantic analysis (LSA), the probabilistic latent semantic analysis (PLSA), Or by a method such as quantification III.

In step S201, the metadata clustering unit 201 may vector the music metadata.

In step S202, the metadata clustering unit 201 clusters the metadata of each piece of music. For example, in step S202, the metadata clustering unit 201 soft clusters the metadata of each piece of music.

More specifically, for example, as shown in FIG. 15, the metadata clustering unit 201 is configured such that the sum of the attribution weights of the clusters of the items in each hierarchy is 1 in each hierarchy. Soft cluster metadata.

For example, the attribution weights to the first cluster, the second cluster, the third cluster, and the fourth cluster in the first hierarchy of the music metadata specified by the music ID of ABC123 are 0.0, 0.8, and 0.0, respectively. , And 0.2. Attribution weights to the fifth cluster, the sixth cluster, the seventh cluster, and the eighth cluster in the second layer of the music metadata specified by the music ID that is ABC123 are 0.4, 0.6, 0.0, and 0.0, respectively. . Attribution weights to the ninth cluster, the tenth cluster, and the eleventh cluster in the third layer of the music metadata specified by the music ID being ABC123 are 0.0, 0.0, and 1.0, respectively. In addition, the attribution weights to each of the four clusters in the nth hierarchy of the music metadata specified by the music ID being ABC123 are 1.0, 0.0, 0.0, and 0.0, respectively.

For example, the attribution weights to the first cluster, the second cluster, the third cluster, and the fourth cluster in the first layer of the music metadata specified by the music ID of CTH863 are 1.0, 0.0, and 0.0, respectively. , And 0.0. Attribution weights to the fifth cluster, sixth cluster, seventh cluster, and eighth cluster in the second layer of the music metadata specified by the music ID being CTH863 are 0.0, 0.5, 0.5, and 0.0, respectively. . Attribution weights to the ninth cluster, the tenth cluster, and the eleventh cluster in the third layer of the music metadata specified by the music ID being CTH863 are 0.7, 0.3, and 0.0, respectively. In addition, the attribution weight to each of the four clusters in the nth hierarchy of the music metadata specified by the music ID which is CTH863 is 0.0, 0.8, 0.2, and 0.0, respectively.

For example, the attribution weights to the first cluster, the second cluster, the third cluster, and the fourth cluster in the first hierarchy of the music metadata specified by the music ID of XYZ567 are 0.0, 0.4, and 0.6, respectively. , And 0.0. Attribution weights to the fifth cluster, the sixth cluster, the seventh cluster, and the eighth cluster in the second hierarchy of the music metadata specified by the music ID of XYZ567 are 0.0, 0.0, 0.0, and 1.0, respectively. . Attribution weights to the ninth cluster, the tenth cluster, and the eleventh cluster in the third layer of the music metadata specified by the music ID of XYZ567 are 0.9, 0.0, and 0.1, respectively. In addition, the attribution weights to each of the four clusters in the nth hierarchy of the music metadata specified by the music ID of XYZ567 are 0.3, 0.0, 0.0, and 0.7, respectively.

In addition, soft clustering of the metadata of each piece of music is not limited to the sum of the attribution weights of the items, that is, the clusters of the pieces of music, in each hierarchy. It is also possible that each item does not belong to any cluster in each hierarchy.

In step S203, the metadata clustering unit 201 assigns the layers of the cluster.

16 and 17, clustering of metadata and allocation of layers of the cluster will be described. 16 is a diagram illustrating an example of metadata. The metadata shown in FIG. 16 is categorical data of either 0 or 1 for simplicity.

Metadata 1, metadata 2, and metadata 3 belong to meta group 1 as a higher-order classification, and metadata 4, metadata 5, and metadata 6 belong to meta-group 2 as a higher-order classification. For example, metadata relating to the artist belongs to meta group 1, and metadata 1 indicates the appearance of the artist, and metadata 2 indicates that the group is a group. For example, metadata relating to the genre belongs to meta group 2, and metadata 4 represents pops, and metadata 5 represents lock.

In the example shown in FIG. 16, the metadata 1 to metadata 6 of the music specified by the music ID of ABC123 are 1, 1, 1, 1, 1, 1, respectively, of the music specified by the music ID of CTH863. Metadata 1 to metadata 6 are 0, 1, 0, 0, 1, 1, respectively, and metadata 1 to metadata 6 of a piece of music specified by a music ID of XYZ567 are respectively 1, 1, 1, 1, 1, 1. The metadata 1 to metadata 6 of the music specified by the music ID of EKF534 are 1, 0, 1, 0, 0, and 1, respectively, and the metadata 1 to metadata of the music specified by the music ID of OPQ385. 6 is 1, 0, 1, 1, 0, 0, respectively.

At this time, the metadata 1 for the music specified by the music ID of ABC123 to the music specified by the music ID of OPQ385 is regarded as a vector. Similarly, each of the metadata 2 to metadata 6 for the music specified by the music ID of ABC123 to the music specified by the music ID of OPQ385 is regarded as a vector. In other words, the value of one metadata for a plurality of pieces of music is regarded as a vector.

Note the distance between these vectors.

In the example shown in FIG. 16, metadata 1, metadata 3, and metadata 4 regarded as a vector have a cluster within Manhattan distance 1, and metadata 2, metadata 5, metadata 6 have a Manhattan distance. It is arranged in other clusters within one.

Therefore, let these clusters be a new layer of metadata. That is, closer metadata is assigned to each layer of the hierarchy.

17 shows an example of metadata clustered in this way and assigned a layer. In the example shown in FIG. 17, metadata 1, metadata 3, and metadata 4 belong to the first layer, and metadata 2, metadata 5, and metadata 6 belong to the second layer.

In this way, each layer is formed of a set of highly correlated metadata, and the clustering of the contents is performed among them, so that the subtle contents cannot be represented by the normal hierarchical division that allows genres, artists, etc. to be hierarchically. The difference can be reflected in the cluster.

Returning to FIG. 14, in step S204, the music clustering unit 202 clusters music pieces for each layer. That is, the music clustering unit 202 classifies each content into one of a plurality of clusters in each of the assigned hierarchies.

Since step S205 and step S206 are the same as that of step S2 and step S3 of FIG. 6, respectively, the description is abbreviate | omitted.

By doing in this way, content can be clustered by reducing data amount and calculation amount, maintaining the detail (degree of the detail of expression) of content representation by metadata.

In addition, as described above, by hierarchizing metadata, contents can be clustered so that subtle differences between contents can be well represented.

Next, the detail of the similar user detection part 203 is demonstrated.

18 is a block diagram illustrating an example of the configuration of the similar user detection unit 203. The similar user detector 203 includes a normalizer 231, a weight calculator 232, and a similarity calculator 233.

The normalization unit 231 normalizes the symbol vector which is an example of the user's preference information. The weight calculation unit 232 calculates the weight for each layer for each of the users from the symbol vector of each normalized user. The similarity calculating unit 233 calculates a similarity indicating the degree of similarity between the user who is going to recommend the piece of music and the preference of another user, from the weights and the symbol vectors for each layer.

Next, the process of detection of the user X with similar preference by the similar user detection unit 203 corresponding to step S21 of FIG. 8 will be described with reference to the flowchart of FIG. 19.

In step S231, the normalization unit 231 normalizes each preference vector of the user.

20 and 21, normalization of the symbol vector will be described. 20 is a diagram showing an example of a symbol vector of each user generated by the symbol vector generation unit 19 and held in the user history information DB 17. As shown in FIG. That is, FIG. 20 shows an example of a symbol vector before normalization.

Among the elements of the symbol vector shown in Fig. 20, the first four elements belong to the first layer, the next four elements belong to the second layer, and the next three elements belong to the third layer. The last four elements belong to the fourth layer.

In the example shown in FIG. 20, the symbol vector of the user specified by the user ID which is U001 is (0.0, 2.8, 0.0, 2.2, 0.4, 0.6, 0.8, 0.0, 0.5, 0.4, 0.4, 0.0, 0.5, 0.4, 0.0). Here, the first four elements, which are 0.0, 2.8, 0.0, and 2.2, respectively, belong to the first layer, and the next four elements, which are 0.4, 0.6, 0.8, and 0.0, respectively, belong to the second layer, and each is 0.5. The next three elements, 0.4, 0.4 belong to the third layer, and the last four elements, 0.0, 0.5, 0.4, and 0.0, respectively, belong to the fourth layer.

In the example shown in FIG. 20, the symbol vector of the user specified by the user ID which is U002 is (0.2, 0.8, 0.5, 0.6, 0.0, 0.5, 0.5, 0.0, 0.7, 0.3, 0.6, 0.0, 0.6, 0.2, 0.0). Here, the first four elements, which are 0.2, 0.8, 0.5, and 0.6, respectively, belong to the first layer, and the next four elements, which are 0.0, 0.5, 0.5, and 0.0, respectively, belong to the second layer, and 0.7, respectively. The next three elements, 0.3 and 0.6, belong to the third layer, and the last four elements, 0.0, 0.6, 0.2 and 0.0, respectively, belong to the fourth layer.

In the example shown in FIG. 20, the symbol vector of the user specified by the user ID of U003 is (0.0, 2.2, 0.1, 1.6, 0.0, 1.0, 2.0, 1.4, 0.0, 1.2, 0.1, 0.3, 0.4, 0.6, 0.7). Here, the first four elements, 0.0, 2.2, 0.1, and 1.6, respectively, belong to the first layer, and the next four elements, each of 0.0, 1.0, 2.0, and 1.4, respectively, belong to the second layer, respectively. The next three elements, 0.0, 1.2, and 0.1, belong to the third layer, and the last four elements, respectively, 0.3, 0.4, 0.6, and 0.7, belong to the fourth layer.

For example, in step S231, the normalization unit 231 normalizes each symbol vector so that the norm in each layer becomes one.

FIG. 21 is a diagram illustrating an example of a symbol vector in which the symbol vector of FIG. 20 is normalized so that the norm in each layer is 1; FIG.

In the example shown in FIG. 21, the normalized symbol vector of the user specified by the user ID of U001 is (0.0, 0.8, 0.0, 0.6, 0.4, 0.6, 0.7, 0.0, 0.7, 0.5, 0.5, 0.0, 0.5, 0.4, 0.0). Here, the first four elements, which are 0.0, 0.8, 0.0, and 0.6, respectively, belong to the first layer, and the next four elements, which are 0.4, 0.6, 0.7, and 0.0, respectively, belong to the second layer, and 0.7, respectively. The next three elements, 0.5, 0.5 belong to the third layer, and the last four elements, 0.0, 0.5, 0.4, and 0.0, respectively, belong to the fourth layer.

In the example shown in FIG. 21, the normalized symbol vector of the user specified by the user ID of U002 is (0.2, 0.7, 0.4, 0.5, 0.0, 0.7, 0.7, 0.0, 0.7, 0.3, 0.6, 0.0, 0.8, 0.3, 0.0). Here, the first four elements, which are 0.2, 0.7, 0.4, and 0.5, respectively, belong to the first layer, and the next four elements, which are 0.0, 0.7, 0.7, and 0.0, respectively, belong to the second layer, and 0.7, respectively. The next three elements, 0.3 and 0.6, belong to the third layer, and the last four elements, 0.0, 0.8, 0.3 and 0.0, respectively, belong to the fourth layer.

In the example shown in FIG. 21, the normalized symbol vector of the user specified by the user ID of U003 is (0.0, 0.8, 0.0, 0.6, 0.0, 0.4, 0.8, 0.5, 0.0, 1.0, 0.1, 0.3, 0.2, 0.2, 0.3). Here, the first four elements, which are 0.0, 0.8, 0.0, and 0.6, respectively, belong to the first layer, and the next four elements, which are 0.0, 0.4, 0.8, and 0.5, respectively, belong to the second layer, and 0.0, respectively. The next three elements, 1.0 and 0.1, belong to the third layer, and the last four elements, respectively, 0.3, 0.2, 0.2 and 0.3, belong to the fourth layer.

Returning to FIG. 19, in step S232, the weight calculation unit 232 calculates the weight for each of the hierarchies of the preference vector of each user. For example, in step S232, the weight calculation part 232 calculates the weight which is the variance of the element which belongs to one layer for each layer.

FIG. 22 is a diagram illustrating an example of weights that are variances of elements belonging to each hierarchy calculated for each user for each hierarchy. In the example shown in FIG. 22, the weight of the 1st layer, the weight of the 2nd layer, the weight of the 3rd layer, and the weight of the 4th layer with respect to the user specified by the user ID which is U001 are 0.17, 0.10, 0.01, and 0.06.

The weight of the first layer, the weight of the second layer, the weight of the third layer, and the weight of the fourth layer are 0.05, 0.17, 0.05, and 0.16, respectively, for the user specified by the user ID of U002. The weight of the first layer, the weight of the second layer, the weight of the third layer, and the weight of the fourth layer are 0.16, 0.10, 0.31, and 0.00, respectively, for the user specified by the user ID of U003.

In step S233, the similarity calculator 233 calculates the similarity of the weighted symbols for each of the users. In step S234, the similar user detection unit 203 detects a user X having a maximum degree of similarity from the user, and the processing ends.

If equation (1) calculates the similarity (sim (u, v)) of users u and v without weighting, user U002 is a user identified by user ID U001. The similarity degree between the user specified by the ID and the user specified by the user ID of U003 is as shown in FIG.

In Equation 1, L is a value indicating the number of layers of the symbol vector, and l is a value specifying the layer of the symbol vector. C (l) represents the entire cluster of symbol vectors, and c is a value for specifying the cluster. h represents the value of the element of a normalized symbol vector.

The elements of the first layer among the elements of the user's preference vector specified by the user ID of U001 and the elements of the first layer of the elements of the user's preference vector specified by the user ID of U002 are multiplied by the corresponding elements. When the multiplied result is integrated, 0.88, which is a value placed on the first layer of the user ID of U002 in FIG. 23, is obtained. Similarly, for the second layer, the third layer, and the fourth layer, elements of the user's preference vector specified by the user ID of U001 and elements of the user's preference vector specified by the user ID of U002 are corresponding elements. When the result of multiplication is multiplied with each other, the result of multiplication is 0.92, 0.97, 0.50, which are values disposed in each of the second, third, and fourth layers of the user ID, which is U002 in FIG.

Finally, the similarity between the preferences of the user specified by the user ID of U001 and the user specified by the user ID of U002 is 0.88, 0.92 determined for each of the first layer, the second layer, the third layer, and the fourth layer. It is set to 3.27 which is a value obtained by adding, 0.97, and 0.50.

Similarly, elements of the first layer among the elements of the user's preference vector specified by the user ID of U001 and elements of the first layer among the elements of the user's preference vector specified by the user ID of U003 correspond to elements corresponding to each other. Multiplying by and multiplying the result, 1.00, which is a value arranged on the first layer of the user ID of U003 in FIG. Similarly, for the second layer, the third layer, and the fourth layer, elements of the user's preference vector specified by the user ID of U001 and elements of the user's preference vector specified by the user ID of U003 correspond to the corresponding elements. When the result of multiplication is multiplied with each other, the result of multiplication is multiplied, and 0.77, 0.57, and 0.15, which are values disposed in each of the second, third, and fourth layers of the user ID of U003 in FIG. 23, are obtained.

Finally, the similarity between the preferences of the user specified by the user ID U001 and the user specified by the user ID U003 is 1.00, 0.77 determined for each of the first layer, the second layer, the third layer, and the fourth layer. It becomes 2.50 which is the value which added, 0.57, and 0.15.

In this way, if the similarity is calculated without weighting, the similarity between the preferences of the user specified by the user ID of U001 and the user identified by the user ID of U002 is specified by the user specified by the user ID of U001 and the user ID of U003. Since the similarity of the user's preference is larger than that of the user, the user specified by the user ID of U002 is detected as the user X having the maximum similarity of the preference.

In contrast, in step S233, the similarity calculator 233 calculates the similarity (sim (u, v)) of the user u and the user v, which are weighted, by Equation (2).

In Equation 2, L is a value indicating the number of layers of the symbol vector, and l is a value specifying the layer of the symbol vector. C (l) represents the entire cluster of symbol vectors, and c is a value for specifying the cluster. h represents the value of the element of a normalized symbol vector. b represents a weight for each of the layers.

FIG. 24 is a diagram showing an example of the similarity between weighted symbols for a user specified by a user ID of U002 and a user specified by a user ID of U003 when the user X is a user specified by a user ID of U001. . In addition, the value shown in FIG. 24 is 100 times the similarity sim (u, v) calculated by Formula (2).

Among the elements of the user's preference vector specified by the user ID of U001, the weight of the first layer of the user specified by the user ID of U001 is multiplied by each of the elements of the first layer, and the user identified by the user ID of U002. If the weights of the first layer of the user specified by the user ID of U002 are multiplied among the elements of the first layer among the elements of the symbol vector of, and are multiplied by the corresponding elements, the result of multiplication is multiplied. 0.72 which is the value arrange | positioned at the 1st layer of user ID which is U002 of 24 is calculated | required.

Similarly, for the second layer, the third layer, and the fourth layer, each of the elements of the user's preference vector specified by the user ID of U001, the first and second layers of the user specified by the user ID of U001 The weights of the third and fourth layers are multiplied, and each of the elements of the user's preference vector specified by the user ID of U002 is multiplied by the weight of the first layer of the user specified by the user ID of U002. When the multiplied elements are multiplied and the multiplied result is integrated, 1.54, 0.03, and 0.48, which are values disposed in each of the second layer, the third layer, and the fourth layer of the user ID of U002 in FIG. 24, are obtained.

Finally, the similarity between the weighted preferences of the user specified by the user ID of U001 and the user identified by the user ID of U002 is obtained for each of the first layer, the second layer, the third layer, and the fourth layer. It becomes 2.76 which is the value which added 0.72, 1.54, 0.03, and 0.48.

Similarly, among the elements of the user's preference vector specified by the user ID of U001, the weight of the first layer of the user specified by the user ID of U001 is multiplied by each of the elements of the first layer, and specified by the user ID of U003. Of the elements of the user's preference vector, each of the elements of the first layer is multiplied by the weight of the first layer of the user specified by the user ID of U003, multiplied by the corresponding elements, and multiplied by the result. 2.74 which is the value arrange | positioned at the 1st layer of the user ID which is U003 of FIG. 24 is calculated | required. Similarly, for the second layer, the third layer, and the fourth layer, each of the elements of the user's preference vector specified by the user ID of U001, the first and second layers of the user specified by the user ID of U001 The weights of the third and fourth layers are multiplied, and each of the elements of the user's preference vector specified by the user ID of U003 is multiplied by the weight of the first layer of the user specified by the user ID of U003, When the corresponding elements are multiplied and the multiplied result is integrated, 0.79, 0.10, and 0.00, which are values disposed in each of the second layer, the third layer, and the fourth layer of the user ID of U003 in FIG. 24, are obtained.

Finally, the similarity degree between the weighted preferences of the user specified by the user ID of U001 and the user identified by the user ID of U003 is obtained for each of the first layer, the second layer, the third layer, and the fourth layer. It becomes 3.64 which is the value which added 2.74, 0.79, 0.10, and 0.00.

As a result, when the similarity is calculated by weighting, the degree of similarity between the user specified by the user ID of U001 and the user specified by the user ID of U003 is specified by the user specified by the user ID of U001 and the user ID of U002. Since it is larger than the similarity of the user's preference, the user specified by the user ID of U003 is detected as the user X of the maximum similarity of the preference.

If the attention is paid to the symbol vector shown in FIG. 20, since the value of each element of the user's preference vector specified by the user ID of U001 is largely changed in the 1st layer compared with the 2nd layer-the 4th layer, Compared with 2nd-4th layer, it is estimated that the value of each element of a 1st layer further relates to the preference of the user specified by the user ID which is U001.

Note the value of each element of the first layer of the user's preference vector specified by the user ID U002 and the user's preference vector specified by the user ID U003. Each element of the first layer of the user's preference vector specified by the user ID U001 than the value of each element of the user's preference vector specified by the user ID U002 is the value of each element of the first layer. Approximate to the value of. Therefore, compared with the user specified by the user ID of U002, the preference of the user specified by the user ID of U003 is predicted to be similar to that of the user specified by the user ID of U001.

In this way, by weighting, the similarity of the sign that the value changes more greatly can be obtained from the value predicted to be more related to the user's preference compared to the value predicted to be less relevant to the user's preference. More accurately, users with similar symbols can be detected.

In addition, although it was demonstrated in step S232 that the weight calculation part 232 calculates the weight which is the variance of the element which belongs to each hierarchy, for example, it is not limited to this, and further, when the variation of an element in a hierarchy is large. What is necessary is just to calculate the weight which becomes a large value, For example, you may calculate entropy H by Formula (3), and calculate the weight which is a value obtained by subtracting entropy H from 1.

In this way, the amount of calculation for selecting appropriate content can be reduced while minimizing the omission of information. In addition, it is possible to present a content that reliably reflects what information the user pays attention to.

In addition, in this specification, the step performed based on a program includes not only the process performed time-series according to the order described, but also the process performed in parallel or separately, even if it does not necessarily process in time series.

The program may be processed by one computer or may be distributed by a plurality of computers. In addition, the program may be transmitted to a remote computer and executed.

In addition, in this specification, a system shows the whole apparatus comprised by several apparatus.

In addition, embodiment of this invention is not limited to embodiment mentioned above, A various change is possible in the range which does not deviate from the summary of this invention.

Claims (11)

An information processing apparatus for selecting and presenting content that satisfies predetermined conditions from a content group to a user, Content classification means for classifying each content constituting the content group into any one of a plurality of first clusters in each of hierarchies according to metadata of the content; Holding means for maintaining a database indicating a correspondence between each content and the first cluster in the hierarchy where each of the respective content is classified; Management means for managing history information on the content of the user; Selection means for specifying a first cluster of interest based on the history information, and selecting content classified in the specific first cluster; Presentation means for presenting the selected content Information processing apparatus comprising a. The method of claim 1, The selection means, Detecting means for detecting a first user and a second user having similar history information; Specifying means for specifying a first cluster, on which content existing on the history information of the second user, which does not exist on the history information of the first user, is classified; Extracting means for extracting content classified in the specified first cluster, And the presenting means presents the content extracted for the first user. The method of claim 1, Generation means for generating preference information indicating the preference of the user in units of the first cluster based on the history information of the user and the database; Grouping means for grouping users based on the preference information, The selection means, Detecting means for detecting a second user belonging to the same group as the first user; Specifying means for specifying a first cluster, on which content existing on the history information of the second user, which does not exist on the history information of the first user, is classified; Extracting means for extracting content classified in the specified first cluster, And the presenting means presents the content extracted for the first user. The method of claim 1, Generation means for generating preference information indicating the preference of the user in units of the first cluster based on the history information of the user and the database; The selection means, Detecting means for detecting a first user and a second user having a similar symbol represented by the preference information; Specifying means for specifying a first cluster to be focused on based on preference information of the first user and preference information of the second user; Extracting means for extracting content classified in the specified first cluster, And the presenting means presents the content extracted for the first user. The method of claim 4, wherein The detection means, Normalization means for normalizing user's preference information, Weight calculation means for calculating weight for each layer from each user's preference information; And similarity calculation means for calculating a similarity indicating a degree of similarity between the first user among the users and the other users among the users, based on the weight for each layer and the preference information. An information processing apparatus for detecting a second user whose taste is similar to the first user from the calculated similarity. The method of claim 1, Generation means for generating preference information indicating the preference of the user in units of the first cluster based on the history information of the user and the database; Grouping means for grouping users based on the preference information, The selection means, Detecting means for detecting a second user belonging to the same group as the first user; Specifying means for specifying a first cluster to be focused on based on preference information of the first user and preference information of the second user; Extracting means for extracting content classified in the specified first cluster, And the presenting means presents the content extracted for the first user. The method of claim 1, Setting means for setting a keyword for each of the first clusters in which the metadata is classified by the content classification means; Creation means for creating a reason text indicating a reason for presentation of the content using the keyword set by the setting means, And the presenting means also presents the reason text. The method of claim 1, Metadata classification means for classifying metadata of the content into any one of a plurality of second clusters, and assigning the hierarchy to a second cluster; The content classification means classifies each content into any one of a plurality of first clusters in each of the assigned hierarchies. The method of claim 1, Generation means for generating preference information indicating the preference of the user in units of the first cluster based on the history information of the user and the database; The selecting means is an information processing apparatus that selects, among the first clusters of all the hierarchies of all the hierarchies, the content classified in the largest number of the first clusters as the first cluster represented by the preference information. In the information processing method of the information processing apparatus which selects the content which meets a predetermined condition from the content group, and presents it to a user, Each content constituting the content group is classified into one of a plurality of clusters in each layer according to the metadata of the content, Maintaining a database representing a correspondence relationship between each content and the cluster in the hierarchy where each of the respective content is classified; Manage history information on the content of the user, On the basis of the history information, a cluster of interest is specified, and contents classified into the specific cluster are selected; Presenting the selected content An information processing method comprising a step. As a program for selecting content that satisfies predetermined conditions from a content group and recommending it to a user, Each content constituting the content group is classified into one of a plurality of clusters in each layer according to the metadata of the content, Maintaining a database representing a correspondence relationship between each content and the cluster in the hierarchy where each of the respective content is classified; Manage history information on the content of the user, On the basis of the history information, a cluster of interest is specified, and contents classified into the specific cluster are selected; Presenting the selected content A program that causes a computer to execute a process including a step.

Images